Evaluation of Improvement Effect of Traditional Chinese Medicine Acupuncture Therapy on Rheumatoid Arthritis Pain Based on Image Recognition Technology

Rheumatoid arthritis (RA) is a chronic, inflammatory, autoimmune disease that occurs in the joints. Its main symptoms include joint swelling, pain, stiffness, redness, swelling, and limitation of movement, which seriously affects the quality of life of patients. Acupuncture and moxibustion in traditional Chinese medicine (TCM) have accumulated rich experience in the treatment of rheumatoid arthritis and have unique advantages [1-4].

Acupuncture, as one of the therapies of traditional Chinese medicine, is also widely used in the treatment of rheumatoid arthritis. By stimulating specific acupoints with needles, it can regulate meridians, qi and blood, and improve inflammation and pain symptoms of joints. Commonly used acupuncture methods include warm acupuncture, cold acupuncture and electroacupuncture [5-8]. The effect of acupuncture on rheumatoid arthritis varies depending on the condition and type of disease. Rheumatoid arthritis belongs to the category of paralysis in Chinese medicine, and Chinese medicine believes that the disease is caused by wind, cold, dampness, heat and other evils and the body’s own deficiency of positive qi, and the basic principle of treatment is to dredge the meridians and activate collaterals, and to alleviate paralysis and pain [9-12]. Acupuncture and moxibustion have the function of dredging the meridians and channels, harmonizing yin and yang, supporting the positive qi in the body and eliminating the evil qi in the body. In treatment, localized acupuncture points are the main points, such as elbow discomfort, you can choose the acupuncture points with obvious pain and numbness, Quchi, Tianjing, Shakuzawa, Shaohai [13-16]. For wrist discomfort, one can choose Ah Yes point, Yang Chi, Waiguan, Yang Xi, and Carpal Bone. For the knee, one can choose Ah Yes point, Blood Sea, Liang Qiu, Knee Eye, Yang Ling Quan [17-19].

In this paper, 80 patients with rheumatoid arthritis who met the diagnostic criteria of Chinese and Western medicine were selected and divided into the acupuncture treatment group based on image recognition technology and the control group of ordinary acupuncture. After starting the treatment, the VAS scores of the two groups were observed, and the rheumatoid factor (RF) and C-reactive protein (CRP) were statistically analyzed to obtain objective evaluation results. A rheumatoid image recognition algorithm based on migration learning and EfficientNet is proposed, and the sensitivity, specificity, positive predictive value, negative predictive value, and the area under the curve value obtained by plotting the characteristic curve of the subjects are calculated to evaluate whether the recognition technique in this paper can effectively recognize the images of patients with rheumatoid joint pain.

2

Clinical observation studies

2.1

Subjects of study

2.1.1

Source of cases

Convenience sampling method was used in this study, and patients who were hospitalized in the rheumatology department of a tertiary-level Chinese medicine hospital in Guangzhou City between July 2022 and December 2022 were selected.

2.1.2

Determination of sample size

Based on the formula, the calculation of the sample size is carried out, the specific formula and algorithm are as follows: (1) $n = Φ^{2} (\sum_{i = 1}^{2} σ_{i}^{2} / k) I {(\sum_{i = 1}^{2} ((μ_{i} - μ)}^{2} / (k - 1))$ Where, n is the number of cases required for each group of samples; σ_i is the standard deviation of each total; μ_i is the mean of each total, μ = Σμ_i / k; k is the number of sample groups to be compared; Φ value is mainly based on α,β,V₁ = k – 1,V₂ = ∞ look up table. The sample content of each group was calculated to be at least 40 cases, totaling 120 patients, and the sample size was increased by 10% in consideration of the possible shedding and loss of visits in the trial, and 132 patients were finally selected for inclusion.

2.1.3

Selection of cases

1)

Chinese medicine diagnostic criteria

The TCM diagnosis (cold-damp paralytic obstruction type) refers to the TCM diagnostic criteria of rheumatoid arthritis in the Guidelines for Clinical Research of New Chinese Medicines in 2002.

Cold-dampness paralytic obstruction: ①Primary symptom: joints are cold and painful and swollen, pain increases with cold and decreases with heat, joints are unfavorable in flexion and extension, morning stiffness, and joints are deformed. ② secondary symptoms: pale mouth, thirst, bad wind and cold, rainy days aggravated, heavy limbs. ③ Tongue and pulse: pale tongue, white moss, tight pulse.

2)

Inclusion Criteria

(1) Patients who meet the western medical diagnostic criteria for rheumatoid arthritis and whose TCM diagnosis is cold-damp paralytic obstruction.

(2) Age above 18 years old and below 75 years old.

(3) Patients who are conscious and able to cooperate with the study.

(4) Voluntarily sign the informed consent to enter the clinical study.

3)

Exclusion Criteria

(1) Late-stage patients: those with stage IV joint function and severe joint deformity.

(2) Patients with other rheumatic immune system diseases, such as systemic lupus erythematosus and dry syndrome.

(3) Combined with serious primary diseases of heart, brain, liver, kidney and blood system.

(4) Patients who are pregnant or breastfeeding, and patients with mental illness.

(5) Those who have done moxibustion in the past and had adverse reactions.

(6) Those who have broken skin, scar, etc. and are not suitable for moxibustion.

4)

Exclusion Criteria

(1) Subjects with 10% or more of their basic personal data or related clinical data missing and cannot be replaced.

(2) Patients who did not comply with the intervention program developed in this study or accepted other experimental programs during the study.

(3) Patients who withdrew in the middle of the study due to various reasons and could not continue to receive this study.

2.2

Research methodology

2.2.1

Grouping methodology

Randomized grouping and control were used to sequentially number 80 patients diagnosed with wind-cold paralytic blockage type RA knee osteoarthritis and aged 30-65 years old, and the numbering from 1-80 was entered into SPSS25.0 software to generate the corresponding random numbers, which were classified into 40 cases in the treatment group (Traditional Chinese Medicine (TCM) Acupuncture based on Image Recognition Technology (TIAT) group), and 40 cases in the control group (Ordinary Needle Acupuncture (OA) group).

2.2.2

Basic treatment

Patients participating in the subject who suffer from basic internal medicine diseases, for different internal medicine diseases, without affecting the results of the study, refer to the Internal Medicine to be disposed of routinely, to ensure that the patient’s indicators are stable before treatment, such as blood pressure, blood glucose, blood lipids are stable.

2.2.3

Selection of acupuncture points

Selected points for the treatment group: bilateral calyx nose, inner knee eye, Liangqiu, Yanglingquan, Blood Sea, Foot Sanli, Yinlingquan, Ququan, Sanyinjiao, knee Yangguan, Ah Yes points.

Selection of acupoints for the control group: the selection of acupoints for the control group was the same as that for the treatment group, and the positioning of the acupoints referred to the “Names and Positioning of Meridian Points” published by the State Administration of Market Supervision and Administration of the People’s Republic of China (SAMSAR) and the National Standardization Administration of China in 2021.

2.2.4

Traditional Chinese Medicine Acupuncture Therapy Based on Image Recognition Technology

1)

Convolutional Neural Networks

In the field of deep learning, a convolutional neural network is one of many artificial neural networks that are primarily designed to process data types that have a regular arrangement of nets, such as two-dimensional images that are neatly aligned in horizontal and vertical coordinates. Convolutional neural networks are inspired by the organization of the visual cortex in animals, and the original intention of designing such networks was to have the networks mimic the adaptive processing features of the visual cortex and learn its spatial hierarchy [20]. CNNs are usually composed of three sub-modules, which are the convolutional layer, the pooling layer, and the fully connected layer.

(1)

Convolutional layer

The convolutional layer essentially acts as a filter, which contains convolutional kernels with different parameters. In neural networks, such parameters are learned by training on a training set, optimized by backpropagation. Each convolutional layer in the processing process, through the sliding window of the previous layer of input images or feature maps for sliding scanning, with the increase in the number of convolutional layers, each subsequent layer of the output feature image, the area is getting smaller and smaller, but the number of channels is getting larger and larger. It is possible to understand it as a local accumulation of images followed by hierarchical clustering. The convolutional layers are shown in Figure 1.

(2)

Activation function

The activation function maps out the features through the function, which removes the redundancy in the data and at the same time enhances the network’s ability to express the nonlinearities, which makes the overall model adaptive to a variety of data and differentiate the outputs, i.e., it has a strong generalization to new data [21]. The activation function is generally after the convolutional layer, and a few commonly used activation functions are listed here, including Sigmoid, Tanh, and ReLU, which are described next.

a) The Sigmoid activation function is the same as that used in the logistic regression classification algorithm, hence the name logistic function. The input of this function is any real value and the output is a value in the range of 0 to 1. As its input gets larger, the output value gets closer to 1.0, while the smaller the input, the closer the output will be to 0.0. The Sigmoid activation function is calculated as follows: (2) $f (x) = \frac{1}{1 + e^{- x}}$

The function is differentiable. That is, the slope of the Sigmoid curve can be found at any two points. However, for the class of neural networks trained using backpropagation, the problem of gradient is the most important, and the Sigmoid function is close to the saturation region, the gradient is relatively flat, and the gradient tends to be close to 0 after the derivation, resulting in the inability to backpropagate for the parameter updating, which makes it impossible to train the neural network. Therefore, one of the disadvantages of using Sigmoid as an activation function is that it is prone to gradient vanishing affecting the network parameter update.Another shortcoming of the Sigmoid function is that its output is not zero-mean, so the signal output by this function to the neurons of the later layers is also not a zero-mean signal resulting in a large jitter in the gradient.

b) The Tanh function is known as the hyperbolic tangent activation function. It is very similar to the Sigmoid activation function and

It even has the same S-shape.

The Tanh activation function is calculated as follows: (3) $f (x) = \frac{e^{x} - e^{- x}}{e^{x} + e^{- x}}$

The function takes any real value as input and outputs a value in the range of -1 to 1. The larger the input value (the more positive), the closer the output will be to 1.0. The larger the value of the input (the more positive), the closer the output value is to 1.0, while the smaller the value of the input (the more negative), the closer the output will be to -1.0. Compared to the Sigmoid function, the function has zero mean as the output, so the practical application of the function is more effective, but it is still not able to solve the problem of gradient disappearance due to the saturation of the function.

c) Given that the Rectified Linear Unit (ReLU) is not only able to effectively overcome the limitations of other activation functions but is also easy to realize, it is one of the activation functions that are currently used more often. The ReLU function is calculated as follows: (4) $f (x) = max (0, x)$

Specifically, it is less susceptible to the vanishing gradient that prevents deep model training.ReLU makes the output of some neurons 0, which can make the neural net sparse and alleviate the overfitting problem. However, another problem easily derived is that the ReLU function forces the value of a neuron to 0, i.e., stops the neuron, which can easily lead to neuronal necrosis (Dead ReLU Problem). The gradient of the neuron that is forced to stop will always be 0, and it will no longer respond to any input data, which directly leads to the corresponding parameters in the neural network will never be updated.

(3)

Pooling layer

The convolutional layer helps remove redundant information and extract features while reducing the image size, while the pooling layer also serves to reduce the size of the input image. Therefore, the pooling layer is generally placed after the convolutional layer to help reduce the image size and remove excessive redundant information while also achieving faster computation. The advantages of the pooling layer are that it can ensure translation and rotation invariance, reduce the parameters of the network, prevent network overfitting, and improve the robustness of the network. The pooling operation requires setting the corresponding kernel and step size.

(4)

Fully Connected Layer

A fully connected layer maps the input features to the target dimension by connecting all those inputs from the previous layer to each activation unit layer in the next layer. A fully connected neural network is made up of multiple fully connected layers that are stacked on top of each other. The core of a fully connected layer is to design a function that maps the dimension of data with a particular input size from R^m to Rⁿ. The core operation of a fully connected layer is the matrix-vector product.

2)

Rheumatoid Arthritis Image Recognition

(1)

Introduction to EfficientNet network

Since its inception, EfficientNet has achieved excellent results in several large image datasets. The backbone network of the model references the MobileNetV2 network, which consists of multiple improved MBConv. The improved MBConv module first performs 1×1 point-by-point convolution on the input and upscales the output channels according to the expansion ratio, then after the Depthwise convolution of ×, it adds the SE module in the SENet, which introduces an attentional mechanism to make the network focus on the more informative channels, and then finally restore the original number of channels using 1×1 convolutional dimensionality reduction [22].

(2)

Transfer Learning

Transfer learning involves transferring knowledge or solving problems from one domain to another that is related but different. The data in the medical field is in the way of patient privacy, labeling requires specialized medical knowledge and other reasons resulting in a small number of data sets, and the use of transfer learning can be a good solution to this problem.

For the identification of rheumatic joint pain disease, migration learning and convolutional neural networks are the current mainstream practices. This paper proposes a convolutional neural network model classification method for migration learning. Using the pre-trained EfficientNet neural network model, migration learning in the ISIC2018 dataset, the migration learning to the weight parameters are fixed, the top layer is modified for training, to achieve the recognition of rheumatoid joint images, rheumatoid joints recognition flowchart is shown in Figure 2.

In this paper, the training adopts a progressive fine-tuning strategy, and the fine-tuning process is shown in Fig. 3, where the first step trains the initialized fully connected layer, and the second step gradually releases the convolutional layer and fine-tunes the trainable layer until the whole network is trained. Finally, the recognition accuracy of oral diseases is used as an indicator of change to select the best depth of the fine-tuned convolutional layer as the final diagnostic network. The improvement over other progressive fine-tuning is that the combination of two optimization algorithms is used for progressive fine-tuning, first using Adam to quickly approximate the optimal value, and then using the SGD algorithm combined with Momentum Momentum for fine-tuning, which speeds up the iteration speed and at the same time prevents the model from entering the local optimal value. By combining the two algorithms, the model is optimized to improve the overall recognition accuracy.

According to the fine-tuning strategy, the EfficientNetB0 underlying network parameters are first frozen and only the top fully connected layer parameters are trained. The optimization algorithm chosen is Adam, which has low memory requirements and runs efficiently, with the expression shown in equation (5) below.

(5)

w_{t} = w_{t - 1} - α \times \frac{{\hat{m}}_{t}}{\sqrt{{\hat{ν}}_{t}} + o}

In the above equation t denotes the number of times, α denotes the learning rate, w is the weight and ò denotes the step size, ${\hat{m}}_{t}$ is the correction for m_t and ${\hat{v}}_{t}$ is the correction for v_t as shown in expressions (6) and (7) (6) ${\hat{m}}_{t} = \frac{m_{t}}{1 - β_{1}^{t}}$ (7) ${\hat{v}}_{t} = \frac{v_{t}}{1 - β_{2}^{t}}$

In the two equations (6)(7) above, β₁ and β₂ are the exponential decay, m_t is the moving average of the gradient, v_t is the squared gradient, and the expressions for m_t and v_t are shown in Eqs. (8) and (9) below: (8) $m_{t} = β_{1} \times m_{t - 1} + (1 - β_{1}) \times g_{t}$ (9) $v_{t} = β_{2} \times v_{t - 1} + (1 - β_{2}) \times g_{t}^{2}$ g_t is that the gradient is a first order derivative and all the above parameters are set as follows: α = 0.001, β₁ = 0.9, β₂ = 0.999, ò = 10⁻⁸. Training with Adam allows the model to iterate quickly and learn the appropriate size of the weights. After the training of the fully connected layer the model convolutional layer is fine-tuned for training, the fine-tuning is done using the SGD algorithm with additional momentum, the expressions are shown in (10)(11) below: (10) $ν_{t + 1} = β ν_{t} + (1 - β) g$ (11) $w_{t + 1} = w_{t} - α \times v_{t + 1}$

The hyperparameter β in the above two equations is used to regulate the decay effect of the gradient, v_t denotes the initial momentum, g is the gradient, α is the learning rate, and w is the weights. The improvement over the normal SGD algorithm is that the momentum v_t is accumulated at the beginning of training to make the training faster. When oscillating around the local optimum, the gradient g is 0, but because of the initial momentum, the trap of the local optimum can be jumped out.

2.3

Observation indicators

1) A pain score (VAS) based on a visual analog scale: out of 10, a higher score indicates more pain. It is an internationally recognized scoring method as a measure of subjective pain perception in subjects. The patient evaluates subjective pain on a 10 cm visual scale, with the 0 end on the left side representing no pain and the 10 end on the right side representing extreme pain, and the tester records the data based on the scale. ① no pain: 0 ≤ pain value ≤ 1; ② light pain: a little pain, but does not affect the general activities, 1 < pain value ≤ 4; ③ medium pain: pain is obvious, still able to move around, 4 < pain value ≤ 7; ④ severe pain: pain is serious, activities are limited, 7 < pain value ≤ 10.

2) Calculate the patient’s Disease Activity Score (DAS28): when the score is ≤2.4, it is low activity; when the score is 2.5-3.8, it is moderate activity; when the score is >3.8, it is high activity.

3) Morning stiffness score: RA subjects were scored according to the duration of morning stiffness.

2.3.1

Information collection and assessment

After obtaining the informed consent of the hospital and the patients, the members of the group collected the general information, scale, and TCM symptomatic integral scale of the patients who met the criteria of Na, Row when they were admitted to the hospital, and needed to pay close attention to the response of the patients and the adverse reaction situation in the intervention process during the intervention, and evaluated the scale and the TCM symptomatic integral scale of the patients of the two groups again after two courses of the intervention, and the data were recorded into the database to do the post The data were entered into the database for later statistical analysis to determine the efficacy.

2.3.2

Statistical analysis

The data were double-checked and entered into SPSS25.0 statistical software for analysis and processing, in which measures that conformed to normal distribution were expressed as mean ± standard deviation $(\bar{x} \pm s)$ , and measures that did not conform to normal distribution were expressed as “median (first quartile, third quartile)”. If the data obeyed normal distribution, paired t-test was used for intra-group comparisons, ANOVA was used for inter-group comparisons, and if there were differences in inter-group comparisons, further two-by-two comparisons were performed. If they did not obey, non-parametric tests were adopted. Count data were expressed as percentages (%); chi-square or rank-sum tests were used based on whether they were ordered or not, respectively. For all the above analyses, α = 0.05 was set, and P < 0.05 was considered statistically significant difference.

2.3.3

Code of Medical Ethics

The study was approved by the Ethics Committee. Before the start of the trial, the purpose of this study, the content of the intervention to be received, the precautions to be taken, and the possible adverse effects were explained to the patients in detail, and their consent was sought and signed on the premise of ensuring informed consent. The confidentiality principle was strictly adhered to during the study, and the data collected from the patients were used solely for statistical analysis and not for any other purpose.

3

Results

3.1

Baseline analysis

3.1.1

General information

Comparison of the general information of the two groups is shown in Table 1: there is no statistically significant difference between the two groups in terms of gender, age, and duration of disease (P>0.05), and they are comparable.

Table 1.

Data of the two groups of participants

Group	Gender		Age (year)	Course (year)
Group	Male	Female	Age (year)	Course (year)
Treatment group(n=40)	23	17	56.1±9.21	12.04±8.12
Control group(n=40)	27	13	51.42±6.99	11.16±8.13
P value	0.845		0.091	0.711

3.1.2

Indicators of clinical symptoms

Comparison of clinical symptom indexes before treatment between the two groups is shown in Table 2: there is no statistically significant difference in VAS score, morning stiffness score and DAS28 score between the two groups (P>0.05), and they are comparable.

Table 2.

Clinical symptoms of the two groups were compared

Group	Number	VAS score	Stiffness score	DAS28 score
Treatment group	40	8.52±2.41	3.42±2.23	6.44±0.95
Control group	40	8.14±2.82	3.62±2.02	6.64±1.12
P value		0.41	0.85	0.51

3.1.3

Routine test indicators

Comparison of routine test indexes before treatment between the two groups of subjects is shown in Table 3: there is no statistical difference between the routine test indexes (RF, CRP, ESR) of the two groups of subjects (P>0.05), and they are comparable.

Table 3.

Comparison of the subjects’ routine test indexes

Group	Number	RF (IU/Mi)	CRF (mg/L)	ESR (mm/1h)
Treatment group	40	212.71±245.14	22.23±24.87	67.37±36.87
Control group	40	195.87±211.41	20.06±22.43	66.57±35.85
P value		0.78	0.68	0.789

3.2

Effectiveness of Traditional Chinese Medicine Acupuncture Therapy Based on Image Recognition Technology

We input the data from the test set into the pre-trained model and count and count the output results, we can get the four-cell table data, as shown in Table 4. Based on the four-cell table data we counted SEN, SPE, LR+, and LR-, which are 82.6%, 77.98%, 3.75, and 0.22, respectively. Its ROC curve is shown in Figure 4, with an AUC of 0.83.

Table 4.

The model data for the test set

Detection value	True value
Detection value	1:Abnormality	0:Normal
1:Abnormality	52	24
0:Normal	11	85

In addition, we obtained the area under the curve (AUC) values of different depth features for rheumatoid joint discrimination in the five levels of 0, 1, 2, 3, and 4 by using Monte Carlo cross-validation of the ROC, and the results are shown in Table 3. Without applying the condition of “stability feature selection”, the prediction model is used to predict the depth features from different layers, and the AUC values obtained are in the range of 0.57-0.71, as shown in Table 5.The mean values of AUC for the five parallel convolutional layers of 0, 1, 2, 3, and 4 are, respectively, 0.57, 0.60, and 0.70, and the mean values of AUC for the five parallel convolutional layers of 0, 1, 2, 3, and 4 are, respectively: 0.57, 0.60, 0.58, 0.64, and 0.59. The highest AUC mean is from the deep features extracted from layer 3, with an AUC mean of 0.64. It can be seen from the table that, among the different methods of statistical pooling performed in each of these five different layers, the best overall classification performance is achieved by using the maximally-pooled deep features from layer 3, with an AUC of 0.71 (95%) CI, 0.70-0.74). It is clear that the multivariate model using this set of deep features is somewhat more accurate in making predictions than the multivariate models from the other four layers, and it is statistically significantly different from the other four layers two by two (p<0.032).

Table 5.

The AUC value of the convolution layer recognition performance

Convolution layer	Characteristic number	Classification performance:AUC
Convolution layer	Characteristic number	Mean pooling	Maximization	Variance pooling	Summation	Mean
Zero layer	61	0.59	0.57	0.54	0.57	0.57
Layer 1	125	0.61	0.55	0.58	0.59	0.60
Layer 2	248	0.64	0.60	0.60	0.55	0.58
Layer 3	498	0.67	0.71	0.65	0.57	0.64
Layer 4	498	0.64	0.57	0.59	0.60	0.59

Next, we explore the effect of “stability feature selection” on the model test results. We apply the model to the best depth feature set (maximum pooling in layer 3). The performance of the model in classifying joint ultrasound pictures is compared using the area under the curve value of the ROC curve, as shown in Fig. 5. Without applying the “stability feature selection”, the AUC of the model was 0.71 (95% CI, 0.69-0.74 (Figure 5a)); On the contrary, the AUC value of the model reached 0.77 (95% CI, 0.71-0.77) with the use of the “Stability Feature Selection” function, which selects the stability of the deep features extracted from layer 3 with the maximum pooling (Fig. 5b). Obviously, we can see that the predictive power of the maximally pooled deep features extracted in layer 3 is better than random guessing without applying “stable feature selection”. After applying the “stability feature selection”, the performance of the model was indeed improved, but there was no statistically significant difference between the two before and after applying the “stability feature selection” condition (p<0.068).

To further investigate the potential of deep features, we calculated the feature importance of each deep feature in layer 3 as shown in Figure 6. Since the model is learned on a linear combination of the selected features, the normalized absolute value of the learned weight for each feature can indicate the importance of that feature in the multivariate model. The importance of all the selected features is summarized and represented by drawing pictures. It can be seen that of all the features in layer 3, only a relatively small number of deep features have a significant importance for rheumatoid joint recognition, while most of them have a very low importance. One of the deep features with the highest importance is located roughly at an index of 140. In order to explore the importance of the features there and their impact on rheumatoid joints, we input the regions of interest of the captured images into the VGG-16 model from both the forward and reverse sides, and observe how high or low the feature importance of each of them is in the feature map with index of 145 at layer 3. However, since the size of the spatial dimension of the feature map in layer 3 is much smaller than that of the original input image and it also contains more complex features extracted, the resultant change in the feature importance of the image, whether it is input from the forward or the reverse direction, is negligible.

3.3

Analysis of post-treatment outcomes

3.3.1

Indicators of clinical symptoms before and after treatment

1)

VAS scores before and after treatment in both groups of subjects

The results before and after treatment are shown in Table 6 and Figure 7:

(1) There was a statistical difference (P<0.01) in the comparison of VAS scores within the two groups before and after treatment, indicating that joint pain improved in both groups after treatment.

(2) There is a statistical difference in the difference of VAS scores between the two groups after treatment (P<0.05), indicating that joint pain improved more significantly in the treatment group than in the control group after treatment.

Table 6.

Two groups of subjects were treated before and after the vas score

Group	Number	Pretreatment	After treatment
Treatment group	40	7.25±2.08	4.43±1.77^**▲▲
Control group	40	6.91±1.92	5.16±1.91^#

Note: Within groups: **P=0.000<0.01, #P=0.035<0.05; Between groups: ▲▲P=0.041<0.05.

2)

Morning stiffness scores before and after treatment in both groups of subjects

The scoring results are shown in Table 7 and Figure 8:

(1) There was a statistical difference between the morning stiffness scores within the two groups before and after treatment when comparing the scores (P<0.01, P<0.05), indicating that morning stiffness improved in both groups after treatment.

(2) The difference in morning stiffness scores between the two groups after treatment was compared with a statistical difference (P<0.05), indicating that morning stiffness improved more significantly in the treatment group than in the control group after treatment.

Table 7.

Two groups of patients were treated before and after the test

Group	Number	Pretreatment	After treatment
Treatment group	40	3.27±3.03	1.42±1.30^**▲▲
Control group	40	3.65±1.83	2.58±1.91^#

Note: Within groups: **P=0.000<0.01, #P=0.041<0.05; Between groups: ▲▲P=0.041<0.05

3)

DAS28 scores before and after treatment in both groups of subjects

The scoring results are shown in Table 8 and Figure 9:

(1) There was a statistical difference between the DAS28 scores within the two groups before and after treatment when compared (P<0.01, P<0.05), indicating that disease activity improved in both groups after treatment.

(2) There is a statistical difference (P<0.05) when the difference in DAS28 scores between the two groups after treatment is compared, indicating that the improvement in disease activity in the treated group is more obvious than that in the control group after treatment.

Table 8.

Post-treatment DAS28 score

Group	Number	Pretreatment	After treatment
Treatment group	40	6.24±1.13	5.23±0.97^**▲▲
Control group	40	6.65±2.03	5.91±1.65^#

Note: Within groups: **P=0.000<0.01, #P=0.01<0.05; Between groups: ▲▲P=0.046<0.05.

3.3.2

Analysis of routine test indicators before and after treatment

1)

RF levels before and after treatment in the two groups of test subjects

Comparison of the trialists before and after treatment is shown in Table 9 and Figure 10:

(1) There was a statistical difference (P<0.05) in the comparison of RF content within the two groups before and after treatment, indicating that RF content improved in both groups after treatment.

(2) The difference in RF content between the two groups after treatment was compared without statistical difference (P>0.05), indicating that the improvement of RF in the two groups after treatment was not significant.

Table 9.

The two groups of patients were treated with rf(IU/mL)

Group	Number	Pretreatment	After treatment
Treatment group	40	211.31±251.13	134.27±158.03^**▲▲
Control group	40	193.87±212.25	155.64±181.41^#

Note: Within groups: **P=0.022<0.05, #P=0.041<0.05; Between groups: P=0.266>0.05.

2)

CRP levels before and after treatment in the two groups of subjects

The results of comparison of CRP content before and after treatment of the subjects are shown in Table 10 and Figure 11:

(1) There was a statistical difference (P<0.05) in the comparison of CRP content within the two groups before and after treatment, indicating that CRP improved in both groups after treatment.

(2) The difference in CRP content between the two groups after treatment was compared without statistical difference (P>0.05), indicating that the improvement of CRP in the two groups after treatment was not significant.

Table 10.

Both groups were treated with CRP(mg/L)

Group	Number	Pretreatment	After treatment
Treatment group	40	22.16±24.08	10.72±11.53^**
Control group	40	20.87±23.25	11.67±16.01^#

Note: Within groups: **P=0.009<0.01, #P=0.035<0.05; Between groups: P=0.555>0.05.

3)

ESR content before and after treatment in both groups of subjects

The ESR content of the subjects before and after treatment is shown in Table 11 and Figure 12:

(1) There was a statistical difference between the ESR content within the two groups before and after treatment (P<0.05), indicating that ESR improved in both groups after treatment.

(2) There was no statistical difference in the difference in ESR content between the two groups after treatment (P>0.05), indicating that the improvement in ESR in both groups after treatment was not significant.

Table 11.

Two groups of patients were treated before and after the esr(mm/1h)

Group	Number	Pretreatment	After treatment
Treatment group	40	67.51±36.71	40.23±29.51^**
Control group	40	66.75±35.87	51.32±30.53^#

Note: Within groups: **P=0.000<0.01, #P=0.005<0.01; Between groups: P=0.065>0.05.

4

Conclusion

In this paper, the research subjects are extracted, and those who meet the research requirements are divided into different groups to start the experiment, and then the method of rheumatic joint image recognition based on transfer learning and Efficientnet neural network is proposed. First, the effectiveness of image recognition technology is studied, and finally, the results of the treatment methods used by the two groups are analyzed. According to the analysis, it was concluded that the SEN, SPE, LR+, LR-, and AUC of the image recognition technique proposed in this paper for the test set data were 82.6%, 77.98%, 3.75, and 0.22 Though the sensitivity and specificity were relatively low, it still proved that the model had undergone effective learning. The physical changes of the patients before and after acupuncture treatment were observed through the test and it was found that the improvement of clinical symptoms and pain reduction in the treatment group were significantly better than that in the control group, and acupuncture was safe and effective in treating rheumatic joint pain.

Lingua:: Inglese

Frequenza di pubblicazione:: 1 volte all'anno
Argomenti della rivista:: Scienze biologiche, Scienze della vita, altro, Matematica, Matematica applicata, Matematica generale, Fisica, Fisica, altro

Feed RSS della rivista

Evaluation of Improvement Effect of Traditional Chinese Medicine Acupuncture Therapy on Rheumatoid Arthritis Pain Based on Image Recognition Technology

Wenbin Hao

Yaya Li

Yang Lan

Pubblicato online: 21 mar 2025

Ricevuto: 10 ott 2024

Accettato: 02 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0641

Parole chiaveImage recognition, Chinese acupuncture, Rheumatoid arthritis, Convolutional neural network

© 2025 Wenbin Hao et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Parole chiave
Image recognition, Chinese acupuncture, Rheumatoid arthritis, Convolutional neural network